DeepSeek Open-Sources Inference Framework, Claims 155 Trillion Tokens Daily

The open-sourcing matters less for the throughput claim than for the Huawei Ascend compatibility, which signals an inference stack designed to run outside Nvidia-only deployments — a quiet hedge against export controls. The U.S. agency directive and Beijing's pushback show the dispute is now political as much as technical, and the memory-chip selloff suggests markets are reading open inference frameworks as a margin risk to the hardware stack. Whether that read holds, or whether cheaper inference expands total demand, is the open question.